Evidence · ANTs

DenoiseImage

DenoiseImage replaces each voxel by a weighted mean of the voxels whose neighbourhoods look alike, with a Rician or Gaussian noise model. LARMORX’s port is a replica of ANTs 2.6.5’s program on ITK 5.4.5. In reproduce it gives ANTs’ one-thread output bit for bit in every compared case of its two records once ANTs uses correctly rounded maths (40 of 40 and 21 of 21). With ANTs as built, 7 of the 40 and none of the 21 match bit for bit; the others differ only where glibc’s expf is not correctly rounded. ANTs’ own result changes with its thread count, in 85 % of a T1w’s voxels at 12 threads; the replica gives ANTs’ one-thread result on any number of threads. Its default, fix, corrects ANTS-43 and ANTS-44.

Ports
  • replica larmorx Apache-2.0: translated from ANTs 2.6.5's and ITK 5.4.5's source, with ITK's AdaptiveDenoising module (Apache-2.0)
Status
replica validated, with its data panel (marked before the statuses' definition)
Records
standard comparison (ANTs 2.6.5), data panel (ANTs 2.6.5) · benchmark ants-denoise · findings ants-denoise
Reference
DenoiseImage · API page ants-denoise

What is compared

DenoiseImage replaces each voxel by a weighted mean of voxels whose neighbourhoods look alike (patch radius 1 and search radius 2 by default), with a Rician or Gaussian noise model. sMRIPrep calls it on every T1w image before the template is built.

The replica runs beside ANTs 2.6.5’s DenoiseImage on the same files with the same arguments. ANTs is the repository’s build (GCC 14.3, with ITK’s AdaptiveDenoising module), run on one thread as built and with correctly rounded maths, and at more threads for its own variability. fMRIPrep’s own ANTs 2.6.2 runs too, in fMRIPrep’s image.

bash
larmorx ants DenoiseImage -i sub-01_T1w.nii.gz -n Rician -s 1 -o sub-01_desc-denoised_T1w.nii.gz
python
import larmorx as lx

# sMRIPrep's call
res = lx.ants.denoise_image("sub-01_T1w.nii.gz", n_threads=8)
res.denoised.to_filename("sub-01_desc-denoised_T1w.nii.gz")

The API page maps every option. How to read what follows: how to read the evidence.

Its port

ANTs and ITK come under the Apache-2.0 licence, as LARMORX does, so the replica is in larmorx itself. It is translated from ANTs’ and ITK’s source, step by step, to give their results bit for bit. A clean room of ANTs’ programs is planned after FreeSurfer (roadmap).

ANTs’ DenoiseImage changes with its thread count (ANTS-42); the replica gives ANTs’ single-thread result on any number of threads. It computes expf correctly rounded, the same on every platform, which is why ANTs as built differs from it in the last bits where glibc’s is not.

The verdicts at a glance

Agreement, replica against the original: ANTs DenoiseImage 2.6.5 (standard comparison and data panel, 72 cases of the two records), by category and verdict
fMRIPrep's calls6
options22
inputs13
errors10
data panel: fMRIPrep's calls17
data panel: other images4

Select a segment to list its cases below.

In reproduce, every case that produces an image gives ANTs’ output at one thread bit for bit: as ANTs is built where glibc’s expf is correctly rounded for every value the run meets, otherwise once the same ANTs binary uses CORE-MATH’s correctly rounded functions. The invalid inputs are refused by both programs. Nothing is open.

In fMRIPrep: denoising a real T1w

The first step of fMRIPrep’s anatomical workflow, with its own arguments, by ANTs at 1, 2, 3, 4, 8 and 12 threads (six times at 8 and three at 12), by fMRIPrep’s own ANTs, and by the replica. ITK’s threads share the image in slabs, four per thread, and write the voxels near each slab’s border in turn: on the coronal slice below, ANTs at 8 threads departs from its one-thread result in bands along the slabs. fMRIPrep’s own run, at 8 threads with its ANTs 2.6.2 build, differs from ANTs 2.6.5 at 8 threads, although the two builds agree bit for bit at one thread. The replica’s default, fix, then corrects ANTS-43, where ITK’s Rician correction factor becomes NaN (the row replica, fix − reproduce). Choose the panels, the plane, the slice and the window; select a point to read every panel’s exact value there; set two images side by side, swipe or blink between them.

Worked example

fMRIPrep's denoising of a real T1w

What this shows: In reproduce, on 1 and 12 threads, the replica gives ANTs’ output at one thread with correctly rounded maths, bit for bit; ANTs as built differs from it only where glibc’s expf is not correctly rounded (5,522 of 5,898,240 voxels, by at most 0.0002). ANTs’ own result moves with its thread count, by up to 271 at 8 threads against one; fMRIPrep’s output, made by its ANTs 2.6.2 at 8 threads, differs from ANTs 2.6.5’s at 8 threads by up to 73, although the two builds agree bit for bit at one thread; so the replica differs from fMRIPrep’s output by up to 271, within ANTs’ own variability between thread counts. The default, fix, corrects ANTS-43 in this call: 437,915 voxels change, by up to 2.81.

The first step of fMRIPrep 25.2.6’s anatomical workflow on OpenNeuro ds000005 sub-01: the raw T1w and the arguments it gave DenoiseImage (sMRIPrep’s Rician model), given to the repository’s ANTs 2.6.5 at 1 to 12 threads (six times at 8 and three at 12; two of the runs at 8 shown), with correctly rounded maths, to fMRIPrep’s own ANTs at one thread, and to the replica in both modes.

Data
OpenNeuro ds000005 (Tom et al., 2007), snapshot 00001, sub-01’s T1w (PDDL-1.0); shown inside fMRIPrep’s brain mask, dilated by 2 voxels
Node
fMRIPrep 25.2.6, anat_fit_wf/anat_template_wf/denoise, run with NSLOTS=8
Command
DenoiseImage -i /data/sub-01/anat/sub-01_T1w.nii.gz -n Rician -o sub-01_T1w_noise_corrected.nii.gz -s 1
Programs
ANTs v2.6.5 (fdce4d2f84) on ITK v5.4.5 (f51594ad88), x86_64-conda-linux-gnu-g++ (conda-forge gcc 14.3.0-20) 14.3.0; ANTs 2.6.2.dev1-gff61a3d (conda-forge) in fMRIPrep’s image; CORE-MATH preloaded where stated; the replica at commit 1dfe56c0dc60
Repeats
the same command by ANTs 2.6.5 at 8 threads 6 times: 1 distinct result; at 12 threads 3 times: 1 distinct result (2026-10-21; earlier runs at 8 threads differed, findings)
ANTS-43
The replica’s default, fix, changes 437,915 of 5,898,240 values, by up to 2.81 (the metrics below)
Grid
160 × 192 × 192 voxels of 1 × 1.33 × 1.33 mm; shown: 151 × 125 × 111 around the brain
Inputds000005 sub-01 T1w, rawint16, as fMRIPrep read it from the dataset
OriginalANTs DenoiseImage 2.6.51 thread, as built · the repository's build (GCC 14.3), glibc's maths
OriginalANTs DenoiseImage 2.6.58 threads · as built (fMRIPrep ran 8)
Replicareplicalarmorx, Apache-2.0 · 1 thread, reproduce · against ANTs DenoiseImage 2.6.5
the original against itself: 1 and 8 threadsANTs DenoiseImage 2.6.5, 8 threads − 1 threadscale ±109
the original against itself: fMRIPrep's run and the repository's buildfMRIPrep's output − ANTs DenoiseImage 2.6.5, 8 threadsscale ±109
the replica against the originalreplica − ANTs DenoiseImage 2.6.5, 1 thread, as builtscale ±109
the replica against itself: fix and reproduce (ANTS-43)replica, fix − reproduce (ANTS-43)scale ±109

Coronal slice 49 of 125; the crosshair where ANTs at 8 threads differs most from one thread in it; ITK's threads split the image into slabs along z, seen here as bands; the differences on one symmetric log scale over six decades below the largest. The interactive viewer needs JavaScript; the metrics below are complete without it.

What separates fMRIPrep's output from the replica's

    • fMRIPrep 25.2.6, as it ranANTs 2.6.2 in its image, 8 threads (NSLOTS=8)
    • ANTs DenoiseImage 2.6.5, 8 threadsrun 6 times: the same bytes every time
    • ANTs DenoiseImage 2.6.5, 1 thread, as builtthe repository's build (GCC 14.3), glibc's maths
    • ANTs DenoiseImage 2.6.2, fMRIPrep's build, 1 threadconda-forge's build, glibc's maths
    • ANTs DenoiseImage 2.6.5, correctly rounded maths, 1 threadthe same binary, CORE-MATH preloaded
    • replica, reproduce1 and 12 threads

Each box holds runs that give the same bytes. Here ANTs 2.6.5 gave 1 distinct result in 6 runs at 8 threads (at 12, 1 distinct result in 3); fMRIPrep’s run at 8 threads, with its own ANTs 2.6.2 build, gave another, although at one thread the two builds give the same bytes: at several threads the result depends on how the threads share the image (ANTS-42). The replica gives ANTs’ one-thread result with correctly rounded maths on any number of threads; its default, fix, then corrects ANTS-43 (the metrics below).

Every comparison, computed in float64 over every voxel of the outputs

Comparison (A − B), verdict and reasonmax |Δ|Values that differRMS of Δmax |Δ| ÷ range1 − NCC
The original against itself: the yardstick
ANTs DenoiseImage 2.6.5, 8 threads − 1 threadreferenceANTs’ result moves with its thread count (ANTS-42). ANTS-422714,816,840 of 5,898,240 (81.7 %)1.420.1890.00011
ANTs DenoiseImage 2.6.5, 8 threads: another run − run 1referenceThe same bytes.identical: 0 of 5,898,240 values differ
ANTs DenoiseImage 2.6.5, 12 threads − 1 threadreferenceANTs’ result moves with its thread count (ANTS-42). ANTS-422715,029,725 of 5,898,240 (85.3 %)1.70.1890.000157
ANTs DenoiseImage 2.6.2 − 2.6.5, 1 thread each, as builtreferenceANTs’ version and build play no part at one thread.identical: 0 of 5,898,240 values differ
ANTs DenoiseImage 2.6.5, 2 threads − 1 threadreferenceANTs’ result moves with its thread count (ANTS-42). ANTS-4296.81,253,922 of 5,898,240 (21.3 %)0.4970.06781.35e−5
ANTs DenoiseImage 2.6.5, 3 threads − 1 threadreferenceANTs’ result moves with its thread count (ANTS-42). ANTS-421612,091,816 of 5,898,240 (35.5 %)0.6490.1132.3e−5
ANTs DenoiseImage 2.6.5, 4 threads − 1 threadreferenceANTs’ result moves with its thread count (ANTS-42). ANTS-422712,919,781 of 5,898,240 (49.5 %)1.040.1895.91e−5
ANTs DenoiseImage 2.6.5, 8 threads: run 3 − run 1referenceThe same bytes.identical: 0 of 5,898,240 values differ
ANTs DenoiseImage 2.6.5, 8 threads: run 4 − run 1referenceThe same bytes.identical: 0 of 5,898,240 values differ
ANTs DenoiseImage 2.6.5, 8 threads: run 5 − run 1referenceThe same bytes.identical: 0 of 5,898,240 values differ
ANTs DenoiseImage 2.6.5, 8 threads: run 6 − run 1referenceThe same bytes.identical: 0 of 5,898,240 values differ
ANTs DenoiseImage 2.6.5, 12 threads: run 2 − run 1referenceThe same bytes.identical: 0 of 5,898,240 values differ
ANTs DenoiseImage 2.6.5, 12 threads: run 3 − run 1referenceThe same bytes.identical: 0 of 5,898,240 values differ
fMRIPrep's output − ANTs DenoiseImage 2.6.5, 8 threadsreferenceANTs’ result moves with its thread count (ANTS-42). ANTS-4272.6163,367 of 5,898,240 (2.77 %)0.1440.05111.13e−6
The replica against the original and fMRIPrep's output
replica − ANTs DenoiseImage 2.6.5, 1 thread, as builtIdentical with correctly rounded mathsOnly where glibc’s expf is not correctly rounded for a value the run meets: with CORE-MATH preloaded, the same ANTs binary gives the replica’s output bit for bit. platform maths0.0001985,522 of 5,898,240 (0.0936 %)3.43e−71.39e−76.42e−18
replica − ANTs DenoiseImage 2.6.5, 1 thread, correctly rounded mathsBit-identicalWith correctly rounded maths, ANTs computes exactly what the replica computes.identical: 0 of 5,898,240 values differ
replica − fMRIPrep's outputWithin the original's own variabilityfMRIPrep’s output is one of ANTs’ results at 8 threads, which move with the threads’ order (the reference rows: ANTs at 8 threads against one thread, and fMRIPrep’s run against the repository’s at 8); the replica’s is ANTs’ one-thread result with correctly rounded maths. ANTS-422714,900,422 of 5,898,240 (83.1 %)1.430.1910.000111
The replica against itself: thread counts, fix and reproduce
replica, 12 threads − 1 threadBit-identicalThe replica gives ANTs’ single-thread result on any number of threads.identical: 0 of 5,898,240 values differ
replica, fix − reproduce (ANTS-43)Deliberate correctionANTS-43, corrected by default with a warning: where ITK’s Rician correction factor overflows to NaN its bias becomes 0; the replica computes the factor in double with scaled Bessel functions. In reproduce, the replica gives ANTs’ result. ANTS-432.81437,915 of 5,898,240 (7.42 %)0.180.001961.2e−6
replica, fix, 12 threads − 1 threadBit-identicalThe same bytes on any number of threads.identical: 0 of 5,898,240 values differ

Δ is A − B. Every number is computed in float64 over all 5,898,240 voxels of the 160 × 192 × 192 outputs, not only the brain shown; the range is B’s maximum minus its minimum. The rows the original against itself are ANTs’ own variability on this input: its thread counts, two runs at 8 threads, its version.

Provenance of this example
  • Data: OpenNeuro ds000005 (Tom et al., 2007), snapshot 00001, sub-01, PDDL-1.0
  • Inputs: T1w (sha256 75c615d7cce397be…); fMRIPrep’s denoised T1w (sha256 ec055df157667867…)
  • Runs: figures cache entry ants-denoise/fmriprep-ds000005-denoise and ants-denoise/fmriprep-ds000005-denoise-repeats, computed 2026-10-21T15:20:52+00:00 with larmorx 1dfe56c0dc60; ANTs v2.6.5 (fdce4d2f84) on ITK v5.4.5 (f51594ad88), x86_64-conda-linux-gnu-g++ (conda-forge gcc 14.3.0-20) 14.3.0; ANTs Version: 2.6.2.dev1-gff61a3d Compiled: Jul 31 2025 06:55:26; every command, its environment, threads and exit status are in its RUN_INFO.json

Every number above is in ants-denoise.json, exported by python -m larmorx_validation evidence ants-denoise.

A deliberate correction: the bias outside a mask (ANTS-44)

With a mask (-x) and the Rician noise model, the patches of the centres in the mask reach the voxels just outside it, which are denoised like the others. ITK smooths and corrects the Rician bias only inside the mask: outside it, those voxels keep the raw noise estimate of the last centre that reached them, an unsmoothed value on the mask’s edge. The input makes sense and the result is wrong, so the replica’s default, fix, gives them the smoothed and corrected bias, as inside; the voxels no centre reached are not denoised, as in ITK; reproduce gives ANTs’ output. fMRIPrep denoises without a mask, so its own call is unchanged; the example is a user’s call, fMRIPrep’s preprocessed T1w of ds000005 denoised within its brain mask.

Worked example

ANTS-44: denoising fMRIPrep's T1w within its brain mask

What this shows: The replica’s default, fix, gives the smoothed and corrected bias to the voxels just outside the mask that the patches reach (ANTS-44), and corrects ANTS-43 inside the mask: 611,567 of the denoised image’s 5,898,240 values change, by up to 177. In reproduce it gives ANTs’ output bit for bit with correctly rounded maths.

DenoiseImage with a mask (-x, Rician noise), on fMRIPrep 25.2.6’s preprocessed T1w of OpenNeuro ds000005 sub-01 and its brain mask. The patches of the centres in the mask reach the voxels just outside it, which are denoised like the others; ITK smooths and corrects the Rician bias only inside the mask and leaves those voxels the raw noise estimate of a centre. The replica’s default, fix, gives them the smoothed and corrected bias, reproduce ANTs’ output.

Data
OpenNeuro ds000005 (mixed-gambles task), v00001, sub-01 (PDDL-1.0)
Command
DenoiseImage -d 3 -i …/out/sub-01/anat/sub-01_desc-preproc_T1w.nii.gz -x …/out/sub-01/anat/sub-01_desc-brain_mask.nii.gz -n Rician -o [denoised.nii.gz,noise.nii.gz]
Defect
ANTS-44: DenoiseImage’s bias outside the mask; corrected by default (decided 2026-10-23), reproduce gives ANTs’ result
Effect
70,886 voxels outside the mask change (the voxels next to it that a centre inside reached), 540,681 inside it (ANTS-43, corrected since 2026-10-12). Outside, the corrected voxels are denoised as the mask’s own edge is: they move from the input by 8.5 on average (ANTs’ 7.0; the mask’s two-voxel inner edge 16.8), and 812 of them become 0 (ANTs’ 3), as the Rician correction makes the darkest voxels inside the mask 0 too (1,299 of the edge’s); Denoised image: 611,567 of 5,898,240 values change, by at most 177; Noise image: 618,279 of 5,898,240 values change, by at most 177
Programs
ANTs v2.6.5 (fdce4d2f84) on ITK v5.4.5 (f51594ad88), x86_64-conda-linux-gnu-g++ (conda-forge gcc 14.3.0-20) 14.3.0, one thread; the replica at commit e22378b480f4, 6 threads
Grid
160 × 192 × 192 voxels of 1 × 1.33 × 1.33 mm; shown: 155 × 129 × 115 around the brain
Inputds000005 sub-01 T1wfMRIPrep's desc-preproc_T1w
OriginalANTs DenoiseImage 2.6.51 thread, as built · the repository's build (GCC 14.3), glibc's maths
Replicareplicalarmorx, Apache-2.0 · fix (the default): ANTS-44 corrected
Replicareplicalarmorx, Apache-2.0 · reproduce: ANTs' behaviour · against ANTs DenoiseImage 2.6.5, correctly rounded maths

Bit-identical5,898,240 of 5,898,240 voxels equal

the replica against the original: correctly rounded mathsreplica, reproduce − ANTs DenoiseImage 2.6.5, correctly rounded maths
the replica against itself: fix and reproduce (ANTS-44)replica, fix − reproduce (ANTS-44)scale ±177

Axial slice 69 of 115; the crosshair where the replica's correction changes the output most outside the mask (ANTS-44; inside it the default also corrects ANTS-43). The interactive viewer needs JavaScript; the metrics below are complete without it.

Every comparison, computed in float64 over every voxel of the outputs

Comparison (A − B), verdict and reasonOutputmax |Δ|Values that differRMS of Δmax |Δ| ÷ range1 − NCC
The original against itself: the yardstick
ANTs DenoiseImage 2.6.5: correctly rounded maths − glibc'sreferenceWhere glibc’s exp, log, sin or cos is not correctly rounded for a value the run meets: ANTs’ own result moves with the maths library.Denoised image0.000183521 of 5,898,240 (0.00883 %)5.79e−76.79e−82.16e−18
Noise image0.000183522 of 5,898,240 (0.00885 %)5.78e−73.11e−72.92e−15
The replica against the original
replica, fix − ANTs DenoiseImage 2.6.5, as builtDeliberate correctionANTS-44, corrected by default with a warning: the voxels outside the mask that the patches of centres in the mask reach are denoised like the others, but ITK smooths and corrects their Rician bias only inside the mask and leaves them the raw noise estimate of the last such centre. The replica gives them the smoothed and corrected bias, as inside; the voxels no centre reached are not denoised, as in ITK. Inside the mask the default also corrects ANTS-43 (the Rician correction factor’s overflow at a high SNR, corrected since 2026-10-12). In reproduce, the replica gives ANTs’ output. ANTS-44Denoised image177611,764 of 5,898,240 (10.4 %)0.9350.06574.98e−6
Noise image177618,472 of 5,898,240 (10.5 %)0.9350.3010.00736
replica, reproduce − ANTs DenoiseImage 2.6.5, as builtIdentical with correctly rounded mathsOnly where glibc’s maths is not correctly rounded: with CORE-MATH preloaded, ANTs gives the replica’s output in reproduce bit for bit (the row against correctly rounded maths).Denoised image0.000183521 of 5,898,240 (0.00883 %)5.79e−76.79e−82.16e−18
Noise image0.000183522 of 5,898,240 (0.00885 %)5.78e−73.11e−72.92e−15
replica, reproduce − ANTs DenoiseImage 2.6.5, correctly rounded mathsBit-identicalIn reproduce, the replica gives ANTs’ result, the defect included.both outputsidentical: 0 of 5,898,240 values differ
The replica against itself: fix and reproduce
replica, fix − reproduce (ANTS-44)Deliberate correctionANTS-44, corrected by default with a warning: the voxels outside the mask that the patches of centres in the mask reach are denoised like the others, but ITK smooths and corrects their Rician bias only inside the mask and leaves them the raw noise estimate of the last such centre. The replica gives them the smoothed and corrected bias, as inside; the voxels no centre reached are not denoised, as in ITK. Inside the mask the default also corrects ANTS-43 (the Rician correction factor’s overflow at a high SNR, corrected since 2026-10-12). In reproduce, the replica gives ANTs’ output. ANTS-44Denoised image177611,567 of 5,898,240 (10.4 %)0.9350.06574.98e−6
Noise image177618,279 of 5,898,240 (10.5 %)0.9350.3010.00736

Δ is A − B, over all 5,898,240 voxels of each output file; the rows the original against itself are ANTs’ own variability on this input.

Provenance of this example
  • Data: OpenNeuro ds000005 (mixed-gambles task), v00001, sub-01, PDDL-1.0
  • Inputs: fMRIPrep’s preprocessed T1w (sha256 8e977ad0c954252d…); fMRIPrep’s brain mask (sha256 726851b7736940c8…)
  • Runs: figures cache entry ants-denoise/ds000005-preproc-masked-ants44, computed 2026-10-23T22:15:22+00:00 with larmorx e22378b480f4; ANTs v2.6.5 (fdce4d2f84) on ITK v5.4.5 (f51594ad88), x86_64-conda-linux-gnu-g++ (conda-forge gcc 14.3.0-20) 14.3.0; every command, its environment, threads and exit status are in its RUN_INFO.json

Every number above is in ants-denoise.json, exported by python -m larmorx_validation evidence ants-denoise.

Against the original’s own variability

Where the replica differs from ANTs as built (glibc’s expf), the record sets the difference against ANTs’ own difference between one thread and 2, 4 and 12 threads.

Variability, the replica against the original divided by the original against itself between thread counts: ANTs DenoiseImage 2.6.5, 32 of the 33 cases that differ from it as built (log scale)
  • the replica's largest |Δ| from ANTs ÷ ANTs' largest |Δ| between its thread counts
  • the replica's (1 − NCC) with ANTs ÷ ANTs' (1 − NCC) between its thread counts
  • inside the original's own variability, ANTs between its thread counts (< 1)
  1. fmriprep/smriprep-t1w denoisedlargest |Δ| 7.3e−7 × the band1 − NCC 0 × the band
  2. fmriprep/smriprep-t1w-ds000210 denoisedlargest |Δ| 3.6e−6 × the band1 − NCC 5e−9 × the band
  3. fmriprep/smriprep-t1w-ds000210 noiselargest |Δ| 3.6e−6 × the band1 − NCC 0 × the band
  4. fmriprep/t1w-preproc-mask denoisedlargest |Δ| 8.1e−7 × the band1 − NCC 4.1e−8 × the band
  5. fmriprep/t1w-preproc-mask noiselargest |Δ| 8.1e−7 × the band1 − NCC 0 × the band
  6. fmriprep/mni-t1w denoisedlargest |Δ| 9.7e−7 × the band1 − NCC 0 × the band
  7. fmriprep/boldref denoisedlargest |Δ| 6e−7 × the band1 − NCC 0 × the band
  8. fmriprep/boldref noiselargest |Δ| 6e−7 × the band1 − NCC 0 × the band
  9. fmriprep/magnitude denoisedlargest |Δ| 3.9e−7 × the band1 − NCC 1.7e−12 × the band
  10. options/defaults denoisedlargest |Δ| 4.8e−6 × the band1 − NCC 5.4e−10 × the band
  11. options/rician denoisedlargest |Δ| 4.2e−7 × the band1 − NCC 0 × the band
  12. options/rician noiselargest |Δ| 4.2e−7 × the band1 − NCC 8.9e−16 × the band
  13. options/gaussian denoisedlargest |Δ| 3.4e−5 × the band1 − NCC 2.6e−8 × the band
  14. options/gaussian noiselargest |Δ| 3.4e−5 × the band1 − NCC 3.7e−9 × the band
  15. options/model-case denoisedlargest |Δ| 4.2e−7 × the band1 − NCC 0 × the band
  16. options/mask-rician denoisedlargest |Δ| 4.4e−7 × the band1 − NCC 1e−13 × the band
  17. options/mask-rician noiselargest |Δ| 4.4e−7 × the band1 − NCC 3.5e−15 × the band
  18. options/mask-gaussian denoisedlargest |Δ| 2.8e−6 × the band1 − NCC 0 × the band
  19. options/mask-fractional denoisedlargest |Δ| 2e−7 × the band1 − NCC 0 × the band
  20. options/mask-values denoisedlargest |Δ| 3.3e−7 × the band1 − NCC 0 × the band
  21. options/mask-larger denoisedlargest |Δ| 4.2e−7 × the band1 − NCC 9.6e−14 × the band
  22. options/missing-mask denoisedlargest |Δ| 4e−7 × the band1 − NCC 0 × the band
  23. options/shrink-2 denoisedlargest |Δ| 3.1e−7 × the band1 − NCC 0 × the band
  24. options/shrink-2 noiselargest |Δ| 2.8e−7 × the band1 − NCC 0 × the band
  25. options/shrink-0 denoisedlargest |Δ| 0.00011 × the band1 − NCC 2e−7 × the band
  26. options/patch-2 denoisedlargest |Δ| 4.1e−7 × the band1 − NCC 2.9e−14 × the band
  27. options/search-1 denoisedlargest |Δ| 5.2e−6 × the band1 − NCC 1.3e−9 × the band
  28. options/search-3 denoisedlargest |Δ| 9.9e−7 × the band1 − NCC 0 × the band
  29. options/radii-per-axis denoisedlargest |Δ| 5.9e−7 × the band1 − NCC 9.5e−14 × the band
  30. options/output-one-parameter denoisedlargest |Δ| 1.5e−6 × the band1 − NCC 6.1e−11 × the band
  31. options/verbose denoisedlargest |Δ| 1.4e−6 × the band1 − NCC 5.6e−11 × the band
  32. inputs/int16-scaled denoisedlargest |Δ| 4.2e−7 × the band1 − NCC 0 × the band
  33. inputs/outliers denoisedlargest |Δ| 9.3e−8 × the band1 − NCC 7.4e−15 × the band
  34. inputs/outliers noiselargest |Δ| 9.3e−8 × the band1 − NCC 0 × the band
  35. inputs/outliers-gaussian denoisedlargest |Δ| 2.1e−6 × the band1 − NCC 0 × the band
  36. inputs/negative-pixdim denoisedlargest |Δ| 1.3e−6 × the band1 − NCC 2.4e−13 × the band
  37. inputs/dim2 denoisedlargest |Δ| 3.6e−7 × the band1 − NCC 0 × the band
  38. inputs/dim2 noiselargest |Δ| 3.6e−7 × the band1 − NCC 2.2e−14 × the band
  39. inputs/dim4 denoisedlargest |Δ| 1.2e−6 × the band1 − NCC 1.6e−13 × the band
  40. inputs/dim4 noiselargest |Δ| 1.2e−6 × the band1 − NCC 1.8e−13 × the band
  41. inputs/dim3-of-4d denoisedlargest |Δ| 4.6e−7 × the band1 − NCC 4.1e−14 × the band
  42. inputs/dimension-from-file denoisedlargest |Δ| 4e−7 × the band1 − NCC 0 × the band
1e-16×1e-15×1e-14×1e-13×1e-12×1e-11×1e-10×1e-09×1e-08×1e-07×1e-06×1e-05×0.0001×0.001×

From the record’s table: ANTs at 2, 4 and 12 threads against 1 thread give ANTs’ thread band for each case and output; a ratio below 1 lies inside it, and a 1 − NCC of 0 is not drawn. Every one of these differences vanishes with correctly rounded maths. Not drawn: options/patch-0 (denoised), where ANTs’ threads change nothing (a band of 0) and the replica differs from ANTs as built by glibc’s maths alone.

On the data panel

The data panel is the fixed set of fourteen OpenNeuro subjects on which LARMORX’s programs are validated: 1.5, 3 and 7 T; Siemens, GE and Philips; single-band, multiband and multi-echo BOLD; a 3.5-year-old to an 82-year-old; a tumour, a stroke with its lesion mask, strong motion. Its record replays every call fMRIPrep 25.2.6 made of the program in its fourteen runs on these subjects, the node’s own command.txt with its inputs from the run’s working directory (calls repeated word for word are compared once), and runs the program on panel images fMRIPrep never gives it. ANTs runs at one thread, as built and with correctly rounded maths, and at 2 and 3 threads for its own variability where it has any; the replica in reproduce, and in fix for the corrections. The panel’s cases are in the case explorer below (the categories data panel: fMRIPrep’s calls and data panel: other images).

21 of 21 bit-identical to ANTs with correctly rounded maths: sMRIPrep’s denoising of every T1w of the fourteen runs (17 calls, the 3.5-year-old, the glioma, the stroke, strong motion, 1.5 T and 7 T among them) and four images fMRIPrep never denoises (the 7 T MP2RAGE, the 82-year-old’s FLAIR, a child’s T2w, the 1.5 T T1w with the Gaussian model). Against ANTs as built, glibc’s expf moves voxels in every case, by at most 0.0015 of ANTs’ own difference between one thread and two or three.

fMRIPrep’s DenoiseImage nodes in the fourteen fMRIPrep runs, and what the panel record compares of them

Node (normalised path)NodesRunsComparedSame call as a compared oneStatusesNot compared: why
anat_fit_wf/anat_template_wf/denoise171417017 pass–

From the panel record (docs/validation/ants-denoise-panel.md): every node of the finished fMRIPrep 25.2.6 runs that ran the program, grouped by its path with the subject, run, space and item removed; a call repeated word for word is compared once.

Images on a small grid (a BOLD reference, a template at 2 mm) are shown as whole volumes; the panel’s large anatomical images (up to 17.5 million voxels) as three full-resolution slices through the voxel where the replica and ANTs as built differ most (the centre of the field of view where they agree), the whole field of view shown, defaced heads included (the data panel’s defacing check).

Worked example

Data panel: sMRIPrep's denoising of the raw 7 T MP2RAGE

What this shows: In reproduce, on 1 and 8 threads, the replica gives ANTs’ bytes with correctly rounded maths; ANTs as built differs only where glibc’s maths is not correctly rounded (largest difference 0.0015). Its default, fix, corrects ANTS-43 here.

sMRIPrep’s DenoiseImage command (-n Rician -s 1, the noise saved) on the raw 7 T MP2RAGE of ds005454 sub-16, the panel’s largest anatomical image (17.5 million voxels, a noisy background).

Data
OpenNeuro ds005454, sub-16 (CC-BY-4.0); the whole field of view shown, only margins where every array is 0 left out
Command
DenoiseImage -i sub-16_acq-mp2rage_T1w.nii.gz -n Rician -o [ denoised.nii.gz, noise.nii.gz ] -s 1
Programs
ANTs v2.6.5 (fdce4d2f84) on ITK v5.4.5 (f51594ad88), x86_64-conda-linux-gnu-g++ (conda-forge gcc 14.3.0-20) 14.3.0; the replica at commit 042d70d5a17c
Grid
224 × 272 × 288 voxels of 0.8 × 0.799 × 0.799 mm; shown 224 × 272 × 288, three full-resolution slices

Data shown

Inputds005454 sub-16sub-16_acq-mp2rage_T1w.nii.gz
OriginalANTs DenoiseImage 2.6.51 thread · as built (GCC 14.3), glibc's maths
Replicareplicalarmorx, Apache-2.0 · reproduce, 8 threads · against ANTs DenoiseImage 2.6.5
Replicareplicalarmorx, Apache-2.0 · fix (the default), 8 threads
the replica against the originalreplica − ANTs DenoiseImage 2.6.5, 1 thread, as builtscale ±20.1

Bit-identical17,547,264 of 17,547,264 voxels equal

the replica against the original: correctly rounded mathsreplica − ANTs DenoiseImage 2.6.5, correctly rounded maths
the replica against itself: fix and reproducereplica, fix − reproducescale ±20.1

Axial slice 171 of 288; the crosshair where the replica differs most from ANTs as built; the three slices through it are stored at full resolution. The interactive viewer needs JavaScript; the metrics below are complete without it.

Stored for this example: three full-resolution slices, one per plane, through display voxel (165, 164, 171): where the main comparison differs most, or the field's centre where nothing differs; every metric is computed over every voxel of the files.

Every comparison, computed in float64 over every voxel of the outputs

Comparison (A − B), verdict and reasonOutputmax |Δ|Values that differRMS of Δmax |Δ| ÷ range1 − NCC
The original against itself: the yardstick
ANTs DenoiseImage 2.6.5, 8 threads − 1 threadreferenceANTs’ own result moves with its thread count (denoised, noise): the yardstick for any difference above.denoised63410,655,570 of 17,547,264 (60.7 %)6.760.1551.79e−5
noise63410,676,569 of 17,547,264 (60.8 %)6.760.1520.000254
The replica against the original
replica − ANTs DenoiseImage 2.6.5, 1 thread, as builtIdentical with correctly rounded mathsOnly where glibc’s maths is not correctly rounded (denoised, noise): with CORE-MATH preloaded, ANTs gives the replica’s bytes.denoised0.001510,951 of 17,547,264 (0.0624 %)6.37e−63.65e−71.59e−17
noise0.001511,205 of 17,547,264 (0.0639 %)6.29e−63.59e−72.2e−16
replica − ANTs DenoiseImage 2.6.5, correctly rounded mathsBit-identicalThe replica gives ANTs’ bytes with correctly rounded maths.both outputsidentical: 0 of 17,547,264 values differ
The replica against itself: thread counts, fix and reproduce
replica, fix − reproduceDeliberate correctionThe replica’s default, fix, corrects ANTS-43 (register); reproduce gives ANTs’ result. Changed: denoised, noise.denoised34.41,858,030 of 17,547,264 (10.6 %)2.020.008411.38e−6
noise34.41,891,826 of 17,547,264 (10.8 %)2.020.008272.12e−5
replica, 1 thread − 8 threads (reproduce)Bit-identicalThe replica’s result does not depend on its thread count.both outputsidentical: 0 of 17,547,264 values differ

Δ is A − B over all 17,547,264 values of each output.

Provenance of this example
  • Data: OpenNeuro ds005454, sub-16, CC-BY-4.0
  • Inputs: input 1 (sha256 971678c9fb76ae12…)
  • Runs: figures cache entry ants-denoise/panel-rician-7t-mp2rage, computed 2026-10-26T23:46:09+00:00 with larmorx 042d70d5a17c; ANTs v2.6.5 (fdce4d2f84) on ITK v5.4.5 (f51594ad88), x86_64-conda-linux-gnu-g++ (conda-forge gcc 14.3.0-20) 14.3.0; every command, its environment, threads and exit status are in its RUN_INFO.json

Every number above is in ants-denoise.json, exported by python -m larmorx_validation evidence ants-denoise.

Worked example

Data panel: fMRIPrep's denoising of a T1w with strong head motion

What this shows: In reproduce, on 1 and 8 threads, the replica gives ANTs’ bytes with correctly rounded maths; ANTs as built differs only where glibc’s maths is not correctly rounded (largest difference 0.00064). Its default, fix, corrects ANTS-43 here.

fMRIPrep 25.2.6’s own call (sMRIPrep’s denoise of each T1w) on ds004173 sub-603309’s MPRAGE acquired with strong instructed head motion (rated 3): its command.txt and input from the fMRIPrep run.

Data
OpenNeuro ds004173, sub-603309 (CC0-1.0); the whole field of view shown, only margins where every array is 0 left out
Node
anat_fit/anat_template/denoise[1] of the fMRIPrep run ds004173-sub-603309-noreconall
Command
DenoiseImage -i sub-603309_acq-headmotion2_T1w.nii.gz -n Rician -o sub-603309_acq-headmotion2_T1w_noise_corrected.nii.gz -s 1
Programs
ANTs v2.6.5 (fdce4d2f84) on ITK v5.4.5 (f51594ad88), x86_64-conda-linux-gnu-g++ (conda-forge gcc 14.3.0-20) 14.3.0; the replica at commit 042d70d5a17c
Grid
192 × 256 × 256 voxels of 1 × 1 × 1 mm; shown 192 × 256 × 256, three full-resolution slices
Inputds004173 sub-603309, MPRAGE with strong motionsub-603309_acq-headmotion2_T1w.nii.gz
OriginalANTs DenoiseImage 2.6.51 thread · as built (GCC 14.3), glibc's maths
Replicareplicalarmorx, Apache-2.0 · reproduce, 8 threads · against ANTs DenoiseImage 2.6.5
Replicareplicalarmorx, Apache-2.0 · fix (the default), 8 threads
the replica against the originalreplica − ANTs DenoiseImage 2.6.5, 1 thread, as builtscale ±4.11

Bit-identical12,582,912 of 12,582,912 voxels equal

the replica against the original: correctly rounded mathsreplica − ANTs DenoiseImage 2.6.5, correctly rounded maths
the replica against itself: fix and reproducereplica, fix − reproducescale ±4.11

Axial slice 83 of 256; the crosshair where the replica differs most from ANTs as built; the three slices through it are stored at full resolution. The interactive viewer needs JavaScript; the metrics below are complete without it.

Stored for this example: three full-resolution slices, one per plane, through display voxel (96, 25, 83): where the main comparison differs most, or the field's centre where nothing differs; every metric is computed over every voxel of the files.

Every comparison, computed in float64 over every voxel of the outputs

Comparison (A − B), verdict and reasonmax |Δ|Values that differRMS of Δmax |Δ| ÷ range1 − NCC
The original against itself: the yardstick
ANTs DenoiseImage 2.6.5, 8 threads − 1 threadreferenceANTs’ own result moves with its thread count (denoised): the yardstick for any difference above.75.78,207,227 of 12,582,912 (65.2 %)0.6890.06281e−5
The replica against the original
replica − ANTs DenoiseImage 2.6.5, 1 thread, as builtIdentical with correctly rounded mathsOnly where glibc’s maths is not correctly rounded (denoised): with CORE-MATH preloaded, ANTs gives the replica’s bytes.0.0006459,043 of 12,582,912 (0.0719 %)5.61e−75.35e−76.64e−18
replica − ANTs DenoiseImage 2.6.5, correctly rounded mathsBit-identicalThe replica gives ANTs’ bytes with correctly rounded maths.identical: 0 of 12,582,912 values differ
The replica against itself: thread counts, fix and reproduce
replica, fix − reproduceDeliberate correctionThe replica’s default, fix, corrects ANTS-43 (register); reproduce gives ANTs’ result. Changed: denoised.4.921,037,637 of 12,582,912 (8.25 %)0.3050.004081.26e−6
replica, 1 thread − 8 threads (reproduce)Bit-identicalThe replica’s result does not depend on its thread count.identical: 0 of 12,582,912 values differ

Δ is A − B over all 12,582,912 values of each output.

Provenance of this example
  • Data: OpenNeuro ds004173, sub-603309, CC0-1.0
  • Inputs: input 1 (sha256 c57cecdd8d2701da…)
  • Runs: figures cache entry ants-denoise/panel-fmriprep-motion, computed 2026-10-27T00:12:13+00:00 with larmorx 042d70d5a17c; ANTs v2.6.5 (fdce4d2f84) on ITK v5.4.5 (f51594ad88), x86_64-conda-linux-gnu-g++ (conda-forge gcc 14.3.0-20) 14.3.0; every command, its environment, threads and exit status are in its RUN_INFO.json

Every number above is in ants-denoise.json, exported by python -m larmorx_validation evidence ants-denoise.

Worked example

Data panel: denoising a FLAIR of 3 mm slices

What this shows: In reproduce, on 1 and 8 threads, the replica gives ANTs’ bytes with correctly rounded maths; ANTs as built differs only where glibc’s maths is not correctly rounded (largest difference 0.00032). Its default, fix, corrects ANTS-43 here.

The same command on an 82-year-old’s FLAIR (ds003592 sub-280: 384 × 512 × 44 voxels of 0.42 × 0.42 × 3 mm).

Data
OpenNeuro ds003592, sub-280 (CC0-1.0); the whole field of view shown, only margins where every array is 0 left out
Command
DenoiseImage -i sub-280_ses-1_FLAIR.nii.gz -n Rician -o [ denoised.nii.gz, noise.nii.gz ] -s 1
Programs
ANTs v2.6.5 (fdce4d2f84) on ITK v5.4.5 (f51594ad88), x86_64-conda-linux-gnu-g++ (conda-forge gcc 14.3.0-20) 14.3.0; the replica at commit 042d70d5a17c
Grid
384 × 512 × 44 voxels of 0.422 × 0.422 × 3 mm; shown 384 × 512 × 44, three full-resolution slices
Inputds003592 sub-280sub-280_ses-1_FLAIR.nii.gz
OriginalANTs DenoiseImage 2.6.51 thread · as built (GCC 14.3), glibc's maths
Replicareplicalarmorx, Apache-2.0 · reproduce, 8 threads · against ANTs DenoiseImage 2.6.5
Replicareplicalarmorx, Apache-2.0 · fix (the default), 8 threads
the replica against the originalreplica − ANTs DenoiseImage 2.6.5, 1 thread, as builtscale ±2

Bit-identical8,650,752 of 8,650,752 voxels equal

the replica against the original: correctly rounded mathsreplica − ANTs DenoiseImage 2.6.5, correctly rounded maths
the replica against itself: fix and reproducereplica, fix − reproducescale ±2

Axial slice 24 of 44; the crosshair where the replica differs most from ANTs as built; the three slices through it are stored at full resolution. The interactive viewer needs JavaScript; the metrics below are complete without it.

Stored for this example: three full-resolution slices, one per plane, through display voxel (180, 41, 24): where the main comparison differs most, or the field's centre where nothing differs; every metric is computed over every voxel of the files.

Every comparison, computed in float64 over every voxel of the outputs

Comparison (A − B), verdict and reasonOutputmax |Δ|Values that differRMS of Δmax |Δ| ÷ range1 − NCC
The original against itself: the yardstick
ANTs DenoiseImage 2.6.5, 8 threads − 1 threadreferenceANTs’ own result moves with its thread count (denoised, noise): the yardstick for any difference above.denoised2137,390,295 of 8,650,752 (85.4 %)2.20.2180.000245
noise2137,406,011 of 8,650,752 (85.6 %)2.20.5050.0437
The replica against the original
replica − ANTs DenoiseImage 2.6.5, 1 thread, as builtIdentical with correctly rounded mathsOnly where glibc’s maths is not correctly rounded (denoised, noise): with CORE-MATH preloaded, ANTs gives the replica’s bytes.denoised0.0003187,787 of 8,650,752 (0.09 %)4.56e−73.25e−71.05e−17
noise0.0003187,913 of 8,650,752 (0.0915 %)4.55e−77.52e−71.85e−15
replica − ANTs DenoiseImage 2.6.5, correctly rounded mathsBit-identicalThe replica gives ANTs’ bytes with correctly rounded maths.both outputsidentical: 0 of 8,650,752 values differ
The replica against itself: thread counts, fix and reproduce
replica, fix − reproduceDeliberate correctionThe replica’s default, fix, corrects ANTS-43 (register); reproduce gives ANTs’ result. Changed: denoised, noise.denoised2.221,050,951 of 8,650,752 (12.1 %)0.2410.002271.98e−6
noise2.221,076,478 of 8,650,752 (12.4 %)0.2410.005260.000463
replica, 1 thread − 8 threads (reproduce)Bit-identicalThe replica’s result does not depend on its thread count.both outputsidentical: 0 of 8,650,752 values differ

Δ is A − B over all 8,650,752 values of each output.

Provenance of this example
  • Data: OpenNeuro ds003592, sub-280, CC0-1.0
  • Inputs: input 1 (sha256 08c1cf34701df2fd…)
  • Runs: figures cache entry ants-denoise/panel-rician-flair, computed 2026-10-27T00:31:00+00:00 with larmorx 042d70d5a17c; ANTs v2.6.5 (fdce4d2f84) on ITK v5.4.5 (f51594ad88), x86_64-conda-linux-gnu-g++ (conda-forge gcc 14.3.0-20) 14.3.0; every command, its environment, threads and exit status are in its RUN_INFO.json

Every number above is in ants-denoise.json, exported by python -m larmorx_validation evidence ants-denoise.

Worked example

Data panel: denoising a child's T2w

What this shows: In reproduce, on 1 and 8 threads, the replica gives ANTs’ bytes with correctly rounded maths; ANTs as built differs only where glibc’s maths is not correctly rounded (largest difference 7.2e-05). Its default, fix, corrects ANTS-43 here.

The same command on a 7.5-year-old’s 0.8 mm T2w (ds000256 sub-CTS203).

Data
OpenNeuro ds000256, sub-CTS203 (PDDL-1.0); the whole field of view shown, only margins where every array is 0 left out
Command
DenoiseImage -i sub-CTS203_T2w.nii.gz -n Rician -o [ denoised.nii.gz, noise.nii.gz ] -s 1
Programs
ANTs v2.6.5 (fdce4d2f84) on ITK v5.4.5 (f51594ad88), x86_64-conda-linux-gnu-g++ (conda-forge gcc 14.3.0-20) 14.3.0; the replica at commit 042d70d5a17c
Grid
224 × 256 × 256 voxels of 0.8 × 0.801 × 0.801 mm; shown 224 × 256 × 256, three full-resolution slices
Inputds000256 sub-CTS203sub-CTS203_T2w.nii.gz
OriginalANTs DenoiseImage 2.6.51 thread · as built (GCC 14.3), glibc's maths
Replicareplicalarmorx, Apache-2.0 · reproduce, 8 threads · against ANTs DenoiseImage 2.6.5
Replicareplicalarmorx, Apache-2.0 · fix (the default), 8 threads
the replica against the originalreplica − ANTs DenoiseImage 2.6.5, 1 thread, as builtscale ±1.41

Bit-identical14,680,064 of 14,680,064 voxels equal

the replica against the original: correctly rounded mathsreplica − ANTs DenoiseImage 2.6.5, correctly rounded maths
the replica against itself: fix and reproducereplica, fix − reproducescale ±1.41

Axial slice 222 of 256; the crosshair where the replica differs most from ANTs as built; the three slices through it are stored at full resolution. The interactive viewer needs JavaScript; the metrics below are complete without it.

Stored for this example: three full-resolution slices, one per plane, through display voxel (111, 193, 222): where the main comparison differs most, or the field's centre where nothing differs; every metric is computed over every voxel of the files.

Every comparison, computed in float64 over every voxel of the outputs

Comparison (A − B), verdict and reasonOutputmax |Δ|Values that differRMS of Δmax |Δ| ÷ range1 − NCC
The original against itself: the yardstick
ANTs DenoiseImage 2.6.5, 8 threads − 1 threadreferenceANTs’ own result moves with its thread count (denoised, noise): the yardstick for any difference above.denoised33.79,734,714 of 14,680,064 (66.3 %)0.2910.1032.08e−5
noise33.79,739,457 of 14,680,064 (66.3 %)0.2910.1360.00102
The replica against the original
replica − ANTs DenoiseImage 2.6.5, 1 thread, as builtIdentical with correctly rounded mathsOnly where glibc’s maths is not correctly rounded (denoised, noise): with CORE-MATH preloaded, ANTs gives the replica’s bytes.denoised7.15e−510,119 of 14,680,064 (0.0689 %)1.92e−72.18e−79.08e−18
noise7.15e−510,285 of 14,680,064 (0.0701 %)1.91e−72.89e−74.42e−16
replica − ANTs DenoiseImage 2.6.5, correctly rounded mathsBit-identicalThe replica gives ANTs’ bytes with correctly rounded maths.both outputsidentical: 0 of 14,680,064 values differ
The replica against itself: thread counts, fix and reproduce
replica, fix − reproduceDeliberate correctionThe replica’s default, fix, corrects ANTS-43 (register); reproduce gives ANTs’ result. Changed: denoised, noise.denoised1.41439,971 of 14,680,064 (3 %)0.07930.004311.36e−6
noise1.41450,361 of 14,680,064 (3.07 %)0.07930.005727.4e−5
replica, 1 thread − 8 threads (reproduce)Bit-identicalThe replica’s result does not depend on its thread count.both outputsidentical: 0 of 14,680,064 values differ

Δ is A − B over all 14,680,064 values of each output.

Provenance of this example
  • Data: OpenNeuro ds000256, sub-CTS203, PDDL-1.0
  • Inputs: input 1 (sha256 ecc1c76bc4c2f1fd…)
  • Runs: figures cache entry ants-denoise/panel-rician-child-t2w, computed 2026-10-27T00:57:52+00:00 with larmorx 042d70d5a17c; ANTs v2.6.5 (fdce4d2f84) on ITK v5.4.5 (f51594ad88), x86_64-conda-linux-gnu-g++ (conda-forge gcc 14.3.0-20) 14.3.0; every command, its environment, threads and exit status are in its RUN_INFO.json

Every number above is in ants-denoise.json, exported by python -m larmorx_validation evidence ants-denoise.

Every case

Each case of the two records, the standard comparison and the data panel, with what it tests, the command line, its verdict and the reason. Search, filter by category or verdict (or select a segment of the chart above), sort by a column; #case=<name> links to one case.

72 cases

CaseCategoryWhat it testsVerdictChecks
fmriprep/smriprep-t1w worked examplefMRIPrep's callssMRIPrep's denoising of each T1w (anatomical.py:1461), as nipype writes it: -i <T1w> -n Rician -o <out> -s 1, no -d; ds000005's raw T1w (160 x 192 x 192)
command lineDenoiseImage -i openneuro/ds000005/sub-01/anat/sub-01_T1w.nii.gz -n Rician -o out.nii.gz -s 1
Identical with correctly rounded maths

denoised: as built, 5522 of 5898240 differ; max |diff| 0.000198, max rel 8.3e-05, 1 - NCC 0 (÷ ANTs’ thread band: 7.3e-07 max |Δ|, 0 1 − NCC); bit-identical with correctly rounded maths

12/12
fmriprep/smriprep-t1w-ds000210fMRIPrep's callsthe same command on ds000210's T1w (256 x 256 x 176, float64 on disk), the noise saved as nipype writes it with save_noise=True: -o [ out, noise ]
command lineDenoiseImage -i openneuro/ds000210/sub-02/anat/sub-02_T1w.nii.gz -n Rician -o [ out.nii.gz, noise.nii.gz ] -s 1
Identical with correctly rounded maths

denoised: as built, 5389 of 11534336 differ; max |diff| 0.00214, max rel 0.00079, 1 - NCC 3.1e-13 (÷ ANTs’ thread band: 3.6e-06 max |Δ|, 5e-09 1 − NCC); noise: as built, 5389 of 11534336 differ; max |diff| 0.00214, max rel 0.011, 1 - NCC 0 (÷ ANTs’ thread band: 3.6e-06 max |Δ|, 0 1 − NCC); bit-identical with correctly rounded maths

20/20
fmriprep/t1w-preproc-maskfMRIPrep's callsfMRIPrep's preprocessed T1w (ds000005) with -n Rician and its brain mask
command lineDenoiseImage -d 3 -i openneuro-derivatives/ds000005-fmriprep/sub-01/anat/sub-01_desc-preproc_T1w.nii.gz -n Rician -x openneuro-derivatives/ds000005-fmriprep/sub-01/anat/sub-01_desc-brain_mask.nii.gz -o [out.nii.gz,noise.nii.gz]
Identical with correctly rounded maths

denoised: as built, 495 of 5898240 differ; max |diff| 0.000122, max rel 7.1e-07, 1 - NCC 1.2e-13 (÷ ANTs’ thread band: 8.1e-07 max |Δ|, 4.1e-08 1 − NCC); noise: as built, 495 of 5898240 differ; max |diff| 0.000122, max rel 0.00056, 1 - NCC 0 (÷ ANTs’ thread band: 8.1e-07 max |Δ|, 0 1 − NCC); bit-identical with correctly rounded maths

20/20
fmriprep/mni-t1wfMRIPrep's callssMRIPrep's options on MNI152NLin2009cAsym's T1w template (res-02, 97 x 115 x 97)
command lineDenoiseImage -i templateflow/tpl-MNI152NLin2009cAsym/tpl-MNI152NLin2009cAsym_res-02_T1w.nii.gz -n Rician -o out.nii.gz -s 1
Identical with correctly rounded maths

denoised: as built, 766 of 1082035 differ; max |diff| 0.000977, max rel 5.9e-07, 1 - NCC 0 (÷ ANTs’ thread band: 9.7e-07 max |Δ|, 0 1 − NCC); bit-identical with correctly rounded maths

12/12
fmriprep/boldreffMRIPrep's callssMRIPrep's options on ds000005's BOLD reference (64 x 64 x 34, int16)
command lineDenoiseImage -i openneuro-derivatives/ds000005-fmriprep/sub-01/func/sub-01_task-mixedgamblestask_run-1_boldref.nii.gz -n Rician -o [out.nii.gz,noise.nii.gz] -s 1
Identical with correctly rounded maths

denoised: as built, 44 of 139264 differ; max |diff| 0.000122, max rel 1.6e-06, 1 - NCC 0 (÷ ANTs’ thread band: 6e-07 max |Δ|, 0 1 − NCC); noise: as built, 44 of 139264 differ; max |diff| 0.000122, max rel 3.7e-05, 1 - NCC 0 (÷ ANTs’ thread band: 6e-07 max |Δ|, 0 1 − NCC); bit-identical with correctly rounded maths

20/20
fmriprep/magnitudefMRIPrep's callssMRIPrep's options on ds003012's fieldmap magnitude (64 x 64 x 36, int32)
command lineDenoiseImage -i openneuro/ds003012/sub-01/fmap/sub-01_magnitude1.nii.gz -n Rician -o out.nii.gz -s 1
Identical with correctly rounded maths

denoised: as built, 112 of 147456 differ; max |diff| 0.00016, max rel 2e-06, 1 - NCC 7.8e-16 (÷ ANTs’ thread band: 3.9e-07 max |Δ|, 1.7e-12 1 − NCC); bit-identical with correctly rounded maths

12/12
options/defaultsoptionsno option but input and output: the Gaussian noise model
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -o out.nii.gz
Identical with correctly rounded maths

denoised: as built, 44 of 35840 differ; max |diff| 3.05e-05, max rel 2.2e-07, 1 - NCC 1.1e-16 (÷ ANTs’ thread band: 4.8e-06 max |Δ|, 5.4e-10 1 − NCC); bit-identical with correctly rounded maths

12/12
options/ricianoptions-n Rician, with the noise image
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -n Rician -o [out.nii.gz,noise.nii.gz]
Identical with correctly rounded maths

denoised: as built, 25 of 35840 differ; max |diff| 3.05e-05, max rel 6.6e-07, 1 - NCC 0 (÷ ANTs’ thread band: 4.2e-07 max |Δ|, 0 1 − NCC); noise: as built, 25 of 35840 differ; max |diff| 3.05e-05, max rel 0.00014, 1 - NCC 1.1e-16 (÷ ANTs’ thread band: 4.2e-07 max |Δ|, 8.9e-16 1 − NCC); bit-identical with correctly rounded maths

20/20
options/gaussianoptions-n Gaussian, with the noise image
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -n Gaussian -o [out.nii.gz,noise.nii.gz]
Identical with correctly rounded maths

denoised: as built, 44 of 35840 differ; max |diff| 3.05e-05, max rel 2.2e-07, 1 - NCC 1.1e-16 (÷ ANTs’ thread band: 3.4e-05 max |Δ|, 2.6e-08 1 − NCC); noise: as built, 45 of 35840 differ; max |diff| 3.05e-05, max rel 3.1e-05, 1 - NCC 5.6e-16 (÷ ANTs’ thread band: 3.4e-05 max |Δ|, 3.7e-09 1 − NCC); bit-identical with correctly rounded maths

20/20
options/model-caseoptions-n RICIAN: the model name is case-insensitive
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -n RICIAN -o out.nii.gz
Identical with correctly rounded maths

denoised: as built, 25 of 35840 differ; max |diff| 3.05e-05, max rel 6.6e-07, 1 - NCC 0 (÷ ANTs’ thread band: 4.2e-07 max |Δ|, 0 1 − NCC); bit-identical with correctly rounded maths

12/12
options/mask-ricianoptions-x a binary mask (uint8), Rician: only voxels in the mask are centres; the bias correction only in the mask
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -n Rician -x '<generated: brain-uint8>' -o [out.nii.gz,noise.nii.gz]
Identical with correctly rounded maths

denoised: as built, 10 of 35840 differ; max |diff| 2.29e-05, max rel 5.2e-07, 1 - NCC 1.1e-16 (÷ ANTs’ thread band: 4.4e-07 max |Δ|, 1e-13 1 − NCC); noise: as built, 10 of 35840 differ; max |diff| 2.29e-05, max rel 7.2e-06, 1 - NCC 2.2e-16 (÷ ANTs’ thread band: 4.4e-07 max |Δ|, 3.5e-15 1 − NCC); bit-identical with correctly rounded maths

20/20
options/mask-gaussianoptions-x a binary mask, Gaussian
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -x '<generated: brain-uint8>' -o out.nii.gz
Identical with correctly rounded maths

denoised: as built, 10 of 35840 differ; max |diff| 1.53e-05, max rel 1.4e-07, 1 - NCC 0 (÷ ANTs’ thread band: 2.8e-06 max |Δ|, 0 1 − NCC); bit-identical with correctly rounded maths

12/12
options/mask-fractionaloptionsa float mask (0, 0.3, 0.6, 0.999, 1, 1.7, 2) read as unsigned char: below 1 is 0
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -n Rician -x '<generated: fractional-float32>' -o out.nii.gz
Identical with correctly rounded maths

denoised: as built, 2 of 35840 differ; max |diff| 7.63e-06, max rel 6.8e-08, 1 - NCC 0 (÷ ANTs’ thread band: 2e-07 max |Δ|, 0 1 − NCC); bit-identical with correctly rounded maths

12/12
options/mask-valuesoptionsa float mask with 0.4, 255, 256, 300.5 and -1: the C++ casts to unsigned char (256 is 0, -1 is 255)
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -n Rician -x '<generated: denoise-mask-values>' -o out.nii.gz
Identical with correctly rounded maths

denoised: as built, 1 of 35840 differ; max |diff| 1.53e-05, max rel 9.6e-08, 1 - NCC 0 (÷ ANTs’ thread band: 3.3e-07 max |Δ|, 0 1 − NCC); bit-identical with correctly rounded maths

12/12
options/mask-largeroptionsa mask larger than the image on the same grid: read by the image's indices
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -n Rician -x '<generated: denoise-mask-larger>' -o out.nii.gz
Identical with correctly rounded maths

denoised: as built, 10 of 35840 differ; max |diff| 2.29e-05, max rel 5.2e-07, 1 - NCC 1.1e-16 (÷ ANTs’ thread band: 4.2e-07 max |Δ|, 9.6e-14 1 − NCC); bit-identical with correctly rounded maths

12/12
options/missing-maskoptionsa mask that does not exist: ANTs prints a message and denoises without a mask
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -n Rician -x none.nii.gz -o out.nii.gz
Identical with correctly rounded maths

denoised: as built, 25 of 35840 differ; max |diff| 3.05e-05, max rel 6.6e-07, 1 - NCC 0 (÷ ANTs’ thread band: 4e-07 max |Δ|, 0 1 − NCC); bit-identical with correctly rounded maths

12/12
options/shrink-2options-s 2: the filter runs on every other voxel; the noise is resampled linearly (float coordinates) onto the full grid
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -n Rician -s 2 -o [out.nii.gz,noise.nii.gz]
Identical with correctly rounded maths

denoised: as built, 93 of 35840 differ; max |diff| 1.91e-05, max rel 2.7e-06, 1 - NCC 0 (÷ ANTs’ thread band: 3.1e-07 max |Δ|, 0 1 − NCC); noise: as built, 160 of 35840 differ; max |diff| 1.72e-05, max rel 4.4e-06, 1 - NCC 0 (÷ ANTs’ thread band: 2.8e-07 max |Δ|, 0 1 − NCC); bit-identical with correctly rounded maths

20/20
options/shrink-3options-s 3, Gaussian
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -s 3 -o [out.nii.gz,noise.nii.gz]
Bit-identical14/14
options/shrink-0options-s 0: ITK makes it 1
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -s 0 -o out.nii.gz
Identical with correctly rounded maths

denoised: as built, 44 of 35840 differ; max |diff| 3.05e-05, max rel 2.2e-07, 1 - NCC 1.1e-16 (÷ ANTs’ thread band: 0.00011 max |Δ|, 2e-07 1 − NCC); bit-identical with correctly rounded maths

12/12
options/shrink-2-maskoptions-s 2 with a mask on the shrunk grid
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -n Rician -s 2 -x '<generated: denoise-mask-shrunk-2>' -o out.nii.gz
Bit-identical9/9
options/patch-2options-p 2: patches of 5 x 5 x 5
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -n Rician -p 2 -o out.nii.gz
Identical with correctly rounded maths

denoised: as built, 25 of 35840 differ; max |diff| 3.05e-05, max rel 8.2e-07, 1 - NCC 5.6e-16 (÷ ANTs’ thread band: 4.1e-07 max |Δ|, 2.9e-14 1 − NCC); bit-identical with correctly rounded maths

12/12
options/patch-0options-p 0: patches of one voxel
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -n Rician -p 0 -o out.nii.gz
Identical with correctly rounded maths

denoised: as built, 11 of 35840 differ; max |diff| 3.05e-05, max rel 8.8e-07, 1 - NCC 1.1e-16 (÷ ANTs’ thread band: inf max |Δ|, inf 1 − NCC); bit-identical with correctly rounded maths

12/12
options/search-1options-r 1
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -r 1 -o out.nii.gz
Identical with correctly rounded maths

denoised: as built, 27 of 35840 differ; max |diff| 2.29e-05, max rel 2.2e-07, 1 - NCC 1.1e-16 (÷ ANTs’ thread band: 5.2e-06 max |Δ|, 1.3e-09 1 − NCC); bit-identical with correctly rounded maths

12/12
options/search-3options-r 3, Rician
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -n Rician -r 3 -o out.nii.gz
Identical with correctly rounded maths

denoised: as built, 23 of 35840 differ; max |diff| 7.44e-05, max rel 3e-06, 1 - NCC 0 (÷ ANTs’ thread band: 9.9e-07 max |Δ|, 0 1 − NCC); bit-identical with correctly rounded maths

12/12
options/radii-per-axisoptions-p 1x2x1 -r 2x1x3: one radius per axis
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -n Rician -p 1x2x1 -r 2x1x3 -o out.nii.gz
Identical with correctly rounded maths

denoised: as built, 29 of 35840 differ; max |diff| 4.2e-05, max rel 1.5e-06, 1 - NCC 6.7e-16 (÷ ANTs’ thread band: 5.9e-07 max |Δ|, 9.5e-14 1 − NCC); bit-identical with correctly rounded maths

12/12
options/output-one-parameteroptions-o [out]: one bracketed name
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -o [out.nii.gz]
Identical with correctly rounded maths

denoised: as built, 44 of 35840 differ; max |diff| 3.05e-05, max rel 2.2e-07, 1 - NCC 1.1e-16 (÷ ANTs’ thread band: 1.5e-06 max |Δ|, 6.1e-11 1 − NCC); bit-identical with correctly rounded maths

12/12
options/no-outputoptionsno -o: the filter runs, nothing is written, exit 0
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -n Rician
Bit-identical1/1
options/verboseoptions-v 1 (stdout is not compared: it holds timings and ITK's object dump)
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -v 1 -o out.nii.gz
Identical with correctly rounded maths

denoised: as built, 44 of 35840 differ; max |diff| 3.05e-05, max rel 2.2e-07, 1 - NCC 1.1e-16 (÷ ANTs’ thread band: 1.4e-06 max |Δ|, 5.6e-11 1 − NCC); bit-identical with correctly rounded maths

12/12
inputs/int16-scaledinputsscaled int16 input
command lineDenoiseImage -d 3 -i '<generated: head-int16-scaled>' -n Rician -o out.nii.gz
Identical with correctly rounded maths

denoised: as built, 21 of 35840 differ; max |diff| 3.05e-05, max rel 2.6e-07, 1 - NCC 0 (÷ ANTs’ thread band: 4.2e-07 max |Δ|, 0 1 − NCC); bit-identical with correctly rounded maths

12/12
inputs/outliersinputshot, negative and zero voxels, a NaN and -inf: they spread to their patches
command lineDenoiseImage -d 3 -i '<generated: head-outliers>' -n Rician -o [out.nii.gz,noise.nii.gz]
Identical with correctly rounded maths

denoised: as built, 21 of 35840 differ; max |diff| 0.0117, max rel 9.1e-07, 1 - NCC 1.7e-15 (÷ ANTs’ thread band: 9.3e-08 max |Δ|, 7.4e-15 1 − NCC); noise: as built, 21 of 35840 differ; max |diff| 0.0117, max rel 4.6e-06, 1 - NCC 0 (÷ ANTs’ thread band: 9.3e-08 max |Δ|, 0 1 − NCC); bit-identical with correctly rounded maths

20/20
inputs/outliers-gaussianinputsthe same, Gaussian
command lineDenoiseImage -d 3 -i '<generated: head-outliers>' -o out.nii.gz
Identical with correctly rounded maths

denoised: as built, 37 of 35840 differ; max |diff| 0.000977, max rel 2.7e-07, 1 - NCC 0 (÷ ANTs’ thread band: 2.1e-06 max |Δ|, 0 1 − NCC); bit-identical with correctly rounded maths

12/12
inputs/constantinputsa constant image (every local variance 0: no centre takes part)
command lineDenoiseImage -d 3 -i '<generated: constant>' -n Rician -o out.nii.gz
Bit-identical9/9
inputs/las-sform-onlyinputsan sform-only LAS image, -s 2 (the noise resampled through the oblique grid)
command lineDenoiseImage -d 3 -i '<generated: las-sform-only>' -n Rician -s 2 -o out.nii.gz
Bit-identical9/9
inputs/negative-pixdiminputsa qform-only image with a negative pixdim
command lineDenoiseImage -d 3 -i '<generated: negative-pixdim-qform>' -n Rician -o out.nii.gz
Identical with correctly rounded maths

denoised: as built, 22 of 35840 differ; max |diff| 9.16e-05, max rel 5.6e-06, 1 - NCC 3.6e-15 (÷ ANTs’ thread band: 1.3e-06 max |Δ|, 2.4e-13 1 − NCC); bit-identical with correctly rounded maths

12/12
inputs/dim2inputs-d 2 on a 2D image, with the noise
command lineDenoiseImage -d 2 -i '<generated: slice-2d>' -n Rician -o [out.nii.gz,noise.nii.gz]
Identical with correctly rounded maths

denoised: as built, 1 of 1280 differ; max |diff| 1.53e-05, max rel 1.1e-07, 1 - NCC 0 (÷ ANTs’ thread band: 3.6e-07 max |Δ|, 0 1 − NCC); noise: as built, 1 of 1280 differ; max |diff| 1.53e-05, max rel 8.8e-07, 1 - NCC 5.6e-16 (÷ ANTs’ thread band: 3.6e-07 max |Δ|, 2.2e-14 1 − NCC); bit-identical with correctly rounded maths

20/20
inputs/dim2-mask-shrinkinputs-d 2 with a 2D mask and -p 2, Gaussian
command lineDenoiseImage -d 2 -i '<generated: slice-2d>' -x '<generated: denoise-slice-mask>' -p 2 -o out.nii.gz
Bit-identical9/9
inputs/dim2-shrink-2inputs-d 2 with -s 2
command lineDenoiseImage -d 2 -i '<generated: slice-2d>' -n Rician -s 2 -o out.nii.gz
Bit-identical9/9
inputs/dim4inputs-d 4: 4D patches and search boxes (time is an axis like the others)
command lineDenoiseImage -d 4 -i '<generated: series-4d>' -n Rician -o [out.nii.gz,noise.nii.gz]
Identical with correctly rounded maths

denoised: as built, 43 of 22400 differ; max |diff| 1.91e-05, max rel 7e-07, 1 - NCC 1.1e-15 (÷ ANTs’ thread band: 1.2e-06 max |Δ|, 1.6e-13 1 − NCC); noise: as built, 46 of 22400 differ; max |diff| 1.91e-05, max rel 2.9e-06, 1 - NCC 1.1e-16 (÷ ANTs’ thread band: 1.2e-06 max |Δ|, 1.8e-13 1 − NCC); bit-identical with correctly rounded maths

20/20
inputs/dim4-shrink-2inputs-d 4 -s 2: time is shrunk too; the noise resampled in 4D
command lineDenoiseImage -d 4 -i '<generated: series-4d>' -s 2 -r 1 -o [out.nii.gz,noise.nii.gz]
Bit-identical14/14
inputs/dim3-of-4dinputs-d 3 on a 4D file: its first volume, with an identity direction
command lineDenoiseImage -d 3 -i '<generated: series-4d>' -n Rician -o out.nii.gz
Identical with correctly rounded maths

denoised: as built, 6 of 4480 differ; max |diff| 1.14e-05, max rel 3.3e-07, 1 - NCC 2.2e-16 (÷ ANTs’ thread band: 4.6e-07 max |Δ|, 4.1e-14 1 − NCC); bit-identical with correctly rounded maths

12/12
inputs/dimension-from-fileinputsno -d: the dimension is the file's
command lineDenoiseImage -i '<generated: head-float32>' -n Rician -o out.nii.gz
Identical with correctly rounded maths

denoised: as built, 25 of 35840 differ; max |diff| 3.05e-05, max rel 6.6e-07, 1 - NCC 0 (÷ ANTs’ thread band: 4e-07 max |Δ|, 0 1 − NCC); bit-identical with correctly rounded maths

12/12
errors/helperrors--help: the usage, exit 0
command lineDenoiseImage --help
Bit-identical1/1
errors/shrink-not-integererrors-s 1.5: the parser's Convert<int> throws (ANTs aborts)
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -s 1.5 -o out.nii.gz
Rejected by both

ANTs exit -6: “s/antsCommandLineParser.h:133:\nITK ERROR: CommandLineParser(0x5d6b64048db0): ERROR: Parse error occured during command line argument processing\nERROR: Unable to convert ‘1.5’ to type ‘i’ as signed int”; the replica exit 1: "larmorx: DenoiseImage:…

1/1
errors/dimension-5errors-d 5: 'Unsupported dimension'
command lineDenoiseImage -d 5 -i '<generated: head-float32>' -o out.nii.gz
Rejected by both

ANTs exit 1: ‘Unsupported dimension’; the replica exit 1: ‘Unsupported dimension’

1/1
errors/missing-inputerrorsan input that does not exist (ANTs crashes)
command lineDenoiseImage -d 3 -i none.nii.gz -o out.nii.gz
Rejected by both

ANTs exit -11: ‘file /none.nii.gz does not exist .’; the replica exit 1: ‘file /none.nii.gz does not exist . \nlarmorx: DenoiseImage: cannot read the input image (ANTs crashes here)’

1/1
errors/unknown-modelerrors-n Poisson: exit 1
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -n Poisson -o out.nii.gz
Rejected by both

ANTs exit 1: ‘’; the replica exit 1: ‘’

1/1
errors/patch-wrong-lengtherrors-p 1x1 in 3D: exit 1
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -p 1x1 -o out.nii.gz
Rejected by both

ANTs exit 1: ‘’; the replica exit 1: ‘’

1/1
errors/search-bracketederrors-r [2]: the option's name is empty, so no radius is read: exit 1
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -r [2] -o out.nii.gz
Rejected by both

ANTs exit 1: ‘’; the replica exit 1: ‘’

1/1
errors/mask-other-griderrorsa mask on another grid: ITK throws, exit 1
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -x '<generated: denoise-mask-other-grid>' -o out.nii.gz
Rejected by both

ANTs exit 1: ‘’; the replica exit 1: ‘’

1/1
errors/mask-full-resolution-shrink-2errors-s 2 with a full-resolution mask: not on the shrunk grid, ITK throws, exit 1
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -s 2 -x '<generated: brain-uint8>' -o out.nii.gz
Rejected by both

ANTs exit 1: ‘’; the replica exit 1: ‘’

1/1
errors/mask-smallererrorsa mask that does not cover the image: ITK's requested region is outside it, exit 1
command lineDenoiseImage -d 3 -i '<generated: head-float32>' -x '<generated: denoise-mask-smaller>' -o out.nii.gz
Rejected by both

ANTs exit 1: ‘’; the replica exit 1: ‘’

1/1
panel/fmriprep/ds000005-sub-01/anat_fit/anat_template/denoise[0]data panel: fMRIPrep's callsfMRIPrep 25.2.6's DenoiseImage node `anat_fit/anat_template/denoise[0]` of the fMRIPrep run ds000005-sub-01 (anat_fit_wf/anat_template_wf/denoise): its command.txt replayed, its inputs from the run's working directory
command lineDenoiseImage -i <testdata-cache>/openneuro/ds000005/sub-01/anat/sub-01_T1w.nii.gz -n Rician -o sub-01_T1w_noise_corrected.nii.gz -s 1
Identical with correctly rounded maths

denoised: as built, 5522 of 5898240 differ; max |diff| 0.000198, max rel 8.3e-05, 1 - NCC 6.4e-18 (÷ ANTs’ thread band: 1.2e-06 max |Δ|, 2.8e-13 1 − NCC); bit-identical with correctly rounded maths

12/12
panel/fmriprep/ds000228-sub-pixar008/anat_fit/anat_template/denoise[0]data panel: fMRIPrep's callsfMRIPrep 25.2.6's DenoiseImage node `anat_fit/anat_template/denoise[0]` of the fMRIPrep run ds000228-sub-pixar008 (anat_fit_wf/anat_template_wf/denoise): its command.txt replayed, its inputs from the run's working directory
command lineDenoiseImage -i <workspace>/oracles/bids-views/ds000228/sub-pixar008/anat/sub-pixar008_T1w.nii.gz -n Rician -o sub-pixar008_T1w_noise_corrected.nii.gz -s 1
Identical with correctly rounded maths

denoised: as built, 3530 of 6488064 differ; max |diff| 0.000132, max rel 2.9e-05, 1 - NCC 5.6e-18 (÷ ANTs’ thread band: – max |Δ|, – 1 − NCC); bit-identical with correctly rounded maths

10/10
panel/fmriprep/ds000256-sub-CTS203/anat_fit/anat_template/denoise[0]data panel: fMRIPrep's callsfMRIPrep 25.2.6's DenoiseImage node `anat_fit/anat_template/denoise[0]` of the fMRIPrep run ds000256-sub-CTS203 (anat_fit_wf/anat_template_wf/denoise): its command.txt replayed, its inputs from the run's working directory
command lineDenoiseImage -i <workspace>/oracles/bids-views/ds000256/sub-CTS203/anat/sub-CTS203_T1w.nii.gz -n Rician -o sub-CTS203_T1w_noise_corrected.nii.gz -s 1
Identical with correctly rounded maths

denoised: as built, 11766 of 14680064 differ; max |diff| 9.92e-05, max rel 9.2e-06, 1 - NCC 8e-18 (÷ ANTs’ thread band: – max |Δ|, – 1 − NCC); bit-identical with correctly rounded maths

10/10
panel/fmriprep/ds001600-sub-1/anat_fit/anat_template/denoise[0]data panel: fMRIPrep's callsfMRIPrep 25.2.6's DenoiseImage node `anat_fit/anat_template/denoise[0]` of the fMRIPrep run ds001600-sub-1 (anat_fit_wf/anat_template_wf/denoise): its command.txt replayed, its inputs from the run's working directory
command lineDenoiseImage -i <testdata-cache>/openneuro/ds001600/sub-1/anat/sub-1_T1w.nii.gz -n Rician -o sub-1_T1w_noise_corrected.nii.gz -s 1
Identical with correctly rounded maths

denoised: as built, 5374 of 11534336 differ; max |diff| 0.000366, max rel 0.00045, 1 - NCC 5.8e-18 (÷ ANTs’ thread band: – max |Δ|, – 1 − NCC); bit-identical with correctly rounded maths

10/10
panel/fmriprep/ds002422-sub-02/anat_fit/anat_template/denoise[0]data panel: fMRIPrep's callsfMRIPrep 25.2.6's DenoiseImage node `anat_fit/anat_template/denoise[0]` of the fMRIPrep run ds002422-sub-02 (anat_fit_wf/anat_template_wf/denoise): its command.txt replayed, its inputs from the run's working directory
command lineDenoiseImage -i <workspace>/oracles/bids-views/ds002422/sub-02/anat/sub-02_T1w.nii -n Rician -o sub-02_T1w_noise_corrected.nii -s 1
Identical with correctly rounded maths

denoised: as built, 9076 of 11534336 differ; max |diff| 0.000793, max rel 0.00026, 1 - NCC 9.2e-18 (÷ ANTs’ thread band: – max |Δ|, – 1 − NCC); bit-identical with correctly rounded maths

10/10
panel/fmriprep/ds002603-sub-01/anat_fit/anat_template/denoise[0]data panel: fMRIPrep's callsfMRIPrep 25.2.6's DenoiseImage node `anat_fit/anat_template/denoise[0]` of the fMRIPrep run ds002603-sub-01 (anat_fit_wf/anat_template_wf/denoise): its command.txt replayed, its inputs from the run's working directory
command lineDenoiseImage -i <workspace>/oracles/bids-views/ds002603/sub-01/anat/sub-01_acq-SaggitalNsig1Mm1_T1w.nii.gz -n Rician -o sub-01_acq-SaggitalNsig1Mm1_T1w_noise_corrected.nii.gz -s 1
Identical with correctly rounded maths

denoised: as built, 4179 of 16384000 differ; max |diff| 0.0288, max rel 2.1e-05, 1 - NCC 5.6e-18 (÷ ANTs’ thread band: – max |Δ|, – 1 − NCC); bit-identical with correctly rounded maths

10/10
panel/fmriprep/ds003592-sub-280/anat_fit/anat_template/denoise[0]data panel: fMRIPrep's callsfMRIPrep 25.2.6's DenoiseImage node `anat_fit/anat_template/denoise[0]` of the fMRIPrep run ds003592-sub-280 (anat_fit_wf/anat_template_wf/denoise): its command.txt replayed, its inputs from the run's working directory
command lineDenoiseImage -i <workspace>/oracles/bids-views/ds003592/sub-280/ses-1/anat/sub-280_ses-1_T1w.nii.gz -n Rician -o sub-280_ses-1_T1w_noise_corrected.nii.gz -s 1
Identical with correctly rounded maths

denoised: as built, 9087 of 12582912 differ; max |diff| 0.000118, max rel 0.001, 1 - NCC 5.6e-18 (÷ ANTs’ thread band: – max |Δ|, – 1 − NCC); bit-identical with correctly rounded maths

10/10
panel/fmriprep/ds004173-sub-603309/anat_fit/anat_template/denoise[0]data panel: fMRIPrep's callsfMRIPrep 25.2.6's DenoiseImage node `anat_fit/anat_template/denoise[0]` of the fMRIPrep run ds004173-sub-603309 (anat_fit_wf/anat_template_wf/denoise): its command.txt replayed, its inputs from the run's working directory
command lineDenoiseImage -i <workspace>/oracles/bids-views/ds004173/sub-603309/anat/sub-603309_acq-headmotion1_T1w.nii.gz -n Rician -o sub-603309_acq-headmotion1_T1w_noise_corrected.nii.gz -s 1
Identical with correctly rounded maths

denoised: as built, 9363 of 12582912 differ; max |diff| 0.000595, max rel 0.043, 1 - NCC 6.1e-18 (÷ ANTs’ thread band: – max |Δ|, – 1 − NCC); bit-identical with correctly rounded maths

10/10
panel/fmriprep/ds004173-sub-603309/anat_fit/anat_template/denoise[1]data panel: fMRIPrep's callsfMRIPrep 25.2.6's DenoiseImage node `anat_fit/anat_template/denoise[1]` of the fMRIPrep run ds004173-sub-603309 (anat_fit_wf/anat_template_wf/denoise): its command.txt replayed, its inputs from the run's working directory
command lineDenoiseImage -i <workspace>/oracles/bids-views/ds004173/sub-603309/anat/sub-603309_acq-headmotion2_T1w.nii.gz -n Rician -o sub-603309_acq-headmotion2_T1w_noise_corrected.nii.gz -s 1
Identical with correctly rounded maths

denoised: as built, 9043 of 12582912 differ; max |diff| 0.000645, max rel 0.00038, 1 - NCC 6.6e-18 (÷ ANTs’ thread band: – max |Δ|, – 1 − NCC); bit-identical with correctly rounded maths

10/10
panel/fmriprep/ds004173-sub-603309/anat_fit/anat_template/denoise[2]data panel: fMRIPrep's callsfMRIPrep 25.2.6's DenoiseImage node `anat_fit/anat_template/denoise[2]` of the fMRIPrep run ds004173-sub-603309 (anat_fit_wf/anat_template_wf/denoise): its command.txt replayed, its inputs from the run's working directory
command lineDenoiseImage -i <workspace>/oracles/bids-views/ds004173/sub-603309/anat/sub-603309_acq-standard_T1w.nii.gz -n Rician -o sub-603309_acq-standard_T1w_noise_corrected.nii.gz -s 1
Identical with correctly rounded maths

denoised: as built, 9224 of 12582912 differ; max |diff| 0.000294, max rel 9.7e-05, 1 - NCC 5.6e-18 (÷ ANTs’ thread band: – max |Δ|, – 1 − NCC); bit-identical with correctly rounded maths

10/10
panel/fmriprep/ds004884-sub-M2079/anat_fit/anat_template/denoise[0]data panel: fMRIPrep's callsfMRIPrep 25.2.6's DenoiseImage node `anat_fit/anat_template/denoise[0]` of the fMRIPrep run ds004884-sub-M2079 (anat_fit_wf/anat_template_wf/denoise): its command.txt replayed, its inputs from the run's working directory
command lineDenoiseImage -i <workspace>/oracles/bids-views/ds004884/sub-M2079/ses-217/anat/sub-M2079_ses-217_acq-tfl3p2_run-3_T1w.nii.gz -n Rician -o sub-M2079_ses-217_acq-tfl3p2_run-3_T1w_noise_corrected.nii.gz -s 1
Identical with correctly rounded maths

denoised: as built, 2260 of 12582912 differ; max |diff| 0.000496, max rel 4.6e-05, 1 - NCC 5.8e-18 (÷ ANTs’ thread band: – max |Δ|, – 1 − NCC); bit-identical with correctly rounded maths

10/10
panel/fmriprep/ds004958-sub-fsm64ps/anat_fit/anat_template/denoise[0]data panel: fMRIPrep's callsfMRIPrep 25.2.6's DenoiseImage node `anat_fit/anat_template/denoise[0]` of the fMRIPrep run ds004958-sub-fsm64ps (anat_fit_wf/anat_template_wf/denoise): its command.txt replayed, its inputs from the run's working directory
command lineDenoiseImage -i <workspace>/oracles/bids-views/ds004958/sub-fsm64ps/anat/sub-fsm64ps_acq-mprageVnav_echo-1_part-mag_T1w.nii.gz -n Rician -o sub-fsm64ps_acq-mprageVnav_echo-1_part-mag_T1w_noise_corrected.nii.gz -s 1
Identical with correctly rounded maths

denoised: as built, 11755 of 11534336 differ; max |diff| 0.000801, max rel 0.00033, 1 - NCC 8e-18 (÷ ANTs’ thread band: – max |Δ|, – 1 − NCC); bit-identical with correctly rounded maths

10/10
panel/fmriprep/ds005003-sub-RS001/anat_fit/anat_template/denoise[0]data panel: fMRIPrep's callsfMRIPrep 25.2.6's DenoiseImage node `anat_fit/anat_template/denoise[0]` of the fMRIPrep run ds005003-sub-RS001 (anat_fit_wf/anat_template_wf/denoise): its command.txt replayed, its inputs from the run's working directory
command lineDenoiseImage -i <workspace>/oracles/bids-views/ds005003/sub-RS001/ses-1/anat/sub-RS001_ses-1_T1w.nii.gz -n Rician -o sub-RS001_ses-1_T1w_noise_corrected.nii.gz -s 1
Identical with correctly rounded maths

denoised: as built, 7785 of 9437184 differ; max |diff| 0.00109, max rel 0.0011, 1 - NCC 9.7e-18 (÷ ANTs’ thread band: – max |Δ|, – 1 − NCC); bit-identical with correctly rounded maths

10/10
panel/fmriprep/ds005040-sub-002/anat_fit/anat_template/denoise[0]data panel: fMRIPrep's callsfMRIPrep 25.2.6's DenoiseImage node `anat_fit/anat_template/denoise[0]` of the fMRIPrep run ds005040-sub-002 (anat_fit_wf/anat_template_wf/denoise): its command.txt replayed, its inputs from the run's working directory
command lineDenoiseImage -i <testdata-cache>/openneuro/ds005040/sub-002/anat/sub-002_T1w.nii.gz -n Rician -o sub-002_T1w_noise_corrected.nii.gz -s 1
Identical with correctly rounded maths

denoised: as built, 5205 of 15842304 differ; max |diff| 0.0159, max rel 9.6e-05, 1 - NCC 5e-18 (÷ ANTs’ thread band: – max |Δ|, – 1 − NCC); bit-identical with correctly rounded maths

10/10
panel/fmriprep/ds005454-sub-16/anat_fit/anat_template/denoise[0]data panel: fMRIPrep's callsfMRIPrep 25.2.6's DenoiseImage node `anat_fit/anat_template/denoise[0]` of the fMRIPrep run ds005454-sub-16 (anat_fit_wf/anat_template_wf/denoise): its command.txt replayed, its inputs from the run's working directory
command lineDenoiseImage -i <testdata-cache>/openneuro/ds005454/sub-16/anat/sub-16_acq-denoised_T1w.nii.gz -n Rician -o sub-16_acq-denoised_T1w_noise_corrected.nii.gz -s 1
Identical with correctly rounded maths

denoised: as built, 8446 of 17547264 differ; max |diff| 0.00394, max rel 0.00026, 1 - NCC 6.3e-18 (÷ ANTs’ thread band: – max |Δ|, – 1 − NCC); bit-identical with correctly rounded maths

10/10
panel/fmriprep/ds005454-sub-16/anat_fit/anat_template/denoise[1]data panel: fMRIPrep's callsfMRIPrep 25.2.6's DenoiseImage node `anat_fit/anat_template/denoise[1]` of the fMRIPrep run ds005454-sub-16 (anat_fit_wf/anat_template_wf/denoise): its command.txt replayed, its inputs from the run's working directory
command lineDenoiseImage -i <testdata-cache>/openneuro/ds005454/sub-16/anat/sub-16_acq-mp2rage_T1w.nii.gz -n Rician -o sub-16_acq-mp2rage_T1w_noise_corrected.nii.gz -s 1
Identical with correctly rounded maths

denoised: as built, 10951 of 17547264 differ; max |diff| 0.0015, max rel 4.8e-05, 1 - NCC 1.6e-17 (÷ ANTs’ thread band: – max |Δ|, – 1 − NCC); bit-identical with correctly rounded maths

10/10
panel/fmriprep/ds005752-sub-ON01016/anat_fit/anat_template/denoise[0]data panel: fMRIPrep's callsfMRIPrep 25.2.6's DenoiseImage node `anat_fit/anat_template/denoise[0]` of the fMRIPrep run ds005752-sub-ON01016 (anat_fit_wf/anat_template_wf/denoise): its command.txt replayed, its inputs from the run's working directory
command lineDenoiseImage -i <testdata-cache>/openneuro/ds005752/sub-ON01016/ses-01/anat/sub-ON01016_ses-01_acq-FSPGR_rec-SCIC_T1w.nii.gz -n Rician -o sub-ON01016_ses-01_acq-FSPGR_rec-SCIC_T1w_noise_corrected.nii.gz -s 1
Identical with correctly rounded maths

denoised: as built, 4286 of 12845056 differ; max |diff| 0.000397, max rel 0.00066, 1 - NCC 5.9e-18 (÷ ANTs’ thread band: – max |Δ|, – 1 − NCC); bit-identical with correctly rounded maths

10/10
panel/images/rician/7t-mp2ragedata panel: other imagessMRIPrep's command (-n Rician -s 1, the noise saved) on the raw 7 T MP2RAGE: the panel's largest anatomical image, its noisy background (ds005454 sub-16: 7 T MP2RAGE uniform image as reconstructed (0.8 mm, 224 x 272 x 288, its noisy background kept), the panel's largest anatomical image)
command lineDenoiseImage -i openneuro/ds005454/sub-16/anat/sub-16_acq-mp2rage_T1w.nii.gz -n Rician -o [ out.nii.gz, noise.nii.gz ] -s 1
Identical with correctly rounded maths

denoised: as built, 10951 of 17547264 differ; max |diff| 0.0015, max rel 4.8e-05, 1 - NCC 1.6e-17 (÷ ANTs’ thread band: 2.4e-06 max |Δ|, 3e-12 1 − NCC); noise: as built, 11205 of 17547264 differ; max |diff| 0.0015, max rel 0.013, 1 - NCC 2.2e-16 (÷ ANTs’ thread band: 2.4e-06 max |Δ|, 2.9e-12 1 − NCC); bit-identical with correctly rounded maths

20/20
panel/images/rician/flairdata panel: other imagesthe same on a FLAIR of 0.42 x 0.42 x 3 mm voxels (384 x 512 x 44) (ds003592 sub-280: an 82-year-old's FLAIR (0.42 x 0.42 x 3 mm, 44 slices))
command lineDenoiseImage -i openneuro/ds003592/sub-280/ses-1/anat/sub-280_ses-1_FLAIR.nii.gz -n Rician -o [ out.nii.gz, noise.nii.gz ] -s 1
Identical with correctly rounded maths

denoised: as built, 7787 of 8650752 differ; max |diff| 0.000318, max rel 0.0069, 1 - NCC 1.1e-17 (÷ ANTs’ thread band: 1.5e-06 max |Δ|, 5.1e-14 1 − NCC); noise: as built, 7913 of 8650752 differ; max |diff| 0.000318, max rel 0.0088, 1 - NCC 1.8e-15 (÷ ANTs’ thread band: 1.5e-06 max |Δ|, 5.1e-14 1 − NCC); bit-identical with correctly rounded maths

20/20
panel/images/rician/child-t2wdata panel: other imagesthe same on a child's 0.8 mm T2w (ds000256 sub-CTS203: a 7.5-year-old's 0.8 mm T2w)
command lineDenoiseImage -i openneuro/ds000256/sub-CTS203/anat/sub-CTS203_T2w.nii.gz -n Rician -o [ out.nii.gz, noise.nii.gz ] -s 1
Identical with correctly rounded maths

denoised: as built, 10119 of 14680064 differ; max |diff| 7.15e-05, max rel 0.0011, 1 - NCC 9.1e-18 (÷ ANTs’ thread band: 3e-06 max |Δ|, 1.5e-12 1 − NCC); noise: as built, 10285 of 14680064 differ; max |diff| 7.15e-05, max rel 0.0029, 1 - NCC 4.4e-16 (÷ ANTs’ thread band: 3e-06 max |Δ|, 1.5e-12 1 − NCC); bit-identical with correctly rounded maths

20/20
panel/images/gaussian/1.5t-float32data panel: other images-n Gaussian on a float32 1.5 T MPRAGE stored uncompressed (ds002422 sub-02: 1.5 T MPRAGE of a 75-year-old (float32, PSR storage, uncompressed))
command lineDenoiseImage -i openneuro/ds002422/sub-02/anat/sub-02_T1w.nii -n Gaussian -o [ out.nii.gz, noise.nii.gz ] -s 1
Identical with correctly rounded maths

denoised: as built, 8150 of 11534336 differ; max |diff| 0.000122, max rel 3.3e-07, 1 - NCC 7.7e-18 (÷ ANTs’ thread band: 0.33 max |Δ|, 0.0065 1 − NCC); noise: as built, 8370 of 11534336 differ; max |diff| 0.000122, max rel 0.0013, 1 - NCC 6.6e-16 (÷ ANTs’ thread band: 0.33 max |Δ|, 0.0065 1 − NCC); bit-identical with correctly rounded maths

20/20

Speed

The wall time of the whole command on six real images, from a BOLD reference and a field map’s magnitude to sMRIPrep’s call on two raw T1w images, with ANTs and the replica at 1 and 12 threads (benchmark).

Speed, the replica against the original: ANTs DenoiseImage 2.6.5

Seconds for the whole command (reading, denoising, writing), the median of repeated runs, on a log scale. The factor is how many times faster the replica is than ANTs, both on one thread.

  • ANTs DenoiseImage, 1 thread
  • ANTs DenoiseImage, 12 threads
  • replica, 1 thread
  • replica, 12 threads
  1. sMRIPrep's call, T1w ds000005ANTs DenoiseImage, 1 thread 217 sANTs DenoiseImage, 12 threads 45.2 sreplica, 1 thread 13 sreplica, 12 threads 2.9 s17×
  2. sMRIPrep's call, T1w ds000210ANTs DenoiseImage, 1 thread 213 sANTs DenoiseImage, 12 threads 46.3 sreplica, 1 thread 16.3 sreplica, 12 threads 3.82 s13×
  3. Rician, MNI template 2 mmANTs DenoiseImage, 1 thread 26 sANTs DenoiseImage, 12 threads 5.69 sreplica, 1 thread 2.44 sreplica, 12 threads 0.55 s11×
  4. Gaussian, T1w ds000005ANTs DenoiseImage, 1 thread 214 sANTs DenoiseImage, 12 threads 45.4 sreplica, 1 thread 12.6 sreplica, 12 threads 2.8 s17×
  5. Rician, BOLD referenceANTs DenoiseImage, 1 thread 1.41 sANTs DenoiseImage, 12 threads 0.282 sreplica, 1 thread 0.206 sreplica, 12 threads 0.0485 s6.8×
  6. Rician, field-map magnitudeANTs DenoiseImage, 1 thread 3.71 sANTs DenoiseImage, 12 threads 0.801 sreplica, 1 thread 0.345 sreplica, 12 threads 0.0748 s11×
0.01 s0.1 s1 s10 s100 s1000 s

From the benchmark report docs/benchmarks/ants-denoise.json (Intel® Core™ i7-8750H CPU @ 2.20GHz, 12 logical CPUs).

Corrections by default

Where ANTs has a deterministic defect that changes a result, the replica corrects it by default, fix, and warns. upstream_defects="reproduce" (on the command line --upstream-defects reproduce from the next release, LARMORX_UPSTREAM_DEFECTS=reproduce in 0.1.0rc1) gives ANTs’ behaviour, and the records compare with it. The record’s list of what fix changed meets ANTS-06, ANTS-43, ANTS-44, ANTS-47, ITK-01 and ITK-02 (register). ANTS-43 changes sMRIPrep’s own command, as the worked example measures. Nondeterministic behaviour, such as the thread dependence (ANTS-42), is never reproduced.

Random testing

Random testing runs random valid inputs and option combinations through the original and the replica, in long runs of several hours: inputs from seeded plain arithmetic (up to 32³ voxels; orientations, data types, scale factors and value ranges of every kind), with options drawn in random combinations from those the program accepts (how random testing works). A difference is shrunk to the smallest input and fewest options that still show it, then fixed and kept as a case of the record, registered as a defect of the original, or reported. The latest long run:

Random testing, the replica against the original: ANTs DenoiseImage 2.6.5, the long run of 2026-10-25 (seed 20261025, 6.2 h for 21 programs): random valid inputs and option combinations, each run through both

ProgramRandom casesExtremeOptions coveredAgreeglibcNoticesUndefinedDifferHeavyGuardUnclear
DenoiseImagereplica1,04909981·26·42···
  • fixed in LARMORX DenoiseImage (replica), 1 case(s): both succeed: DenoiseImage without -d on a 4D image: ANTs takes it as 3D (ANTS-99); LARMORX failed on the file’s dimension. Now reproduce takes 3, the default the file’s dimension, both naming ANTS-99
  • fixed in LARMORX DenoiseImage (replica), 8 case(s): denoised: NCC with ANTs as built, denoised: bit-identical to ANTs with correctly rounded math: as above: the Rician bias’s smoothing kernel on micrometre voxels (ANTS-14 now reproduced and named)
  • fixed in LARMORX DenoiseImage (replica), 6 case(s): denoised: NCC with ANTs as built, denoised: bit-identical to ANTs with correctly rounded math, noise: NCC with ANTs as built, noise: bit-identical to ANTs with correctly rounded math: the Rician bias’s smoothing kernel on micrometre voxels (ANTS-14 now reproduced and named)
  • fixed in LARMORX DenoiseImage (replica), 3 case(s): denoised: affine, denoised: data type and shape, denoised: header (every field): DenoiseImage without -d on a 4D image: ANTs takes it as 3D (ANTS-99); now reproduced, both modes naming it
  • fixed in LARMORX DenoiseImage (replica), 4 case(s): denoised: affine, denoised: data type and shape, denoised: header (every field), noise: affine, noise: data type and shape, noise: header (every field): DenoiseImage without -d on a 4D image: ANTs denoises and writes it as 3D (ANTS-99, as above)
  • fixed in LARMORX DenoiseImage (replica), 10 case(s): denoised: bit-identical to ANTs with correctly rounded math: as above: the Rician bias’s smoothing kernel on micrometre voxels (ANTS-14 now reproduced and named)
  • fixed in LARMORX DenoiseImage (replica), 1 case(s): denoised: bit-identical to ANTs with correctly rounded math, noise: NCC with ANTs as built, noise: bit-identical to ANTs with correctly rounded math: the Rician bias’s smoothing kernel on micrometre voxels (ANTS-14 now reproduced and named)
  • fixed in LARMORX DenoiseImage (replica), 3 case(s): denoised: bit-identical to ANTs with correctly rounded math, noise: bit-identical to ANTs with correctly rounded math: Rician denoising on micrometre voxels: the bias’s smoothing variance of 2 mm² is 2·10⁶ voxels², whose ITK kernel is NaN (ANTS-14), so ITK leaves the bias unsmoothed; LARMORX smoothed it with the corrected kernel in both modes. Now reproduce keeps ITK’s kernel, and both modes name ANTS-14 (larmorx_image::adaptive_denoising)
  • fixed in LARMORX DenoiseImage (replica), 5 case(s): denoised: data type and shape, denoised: header (every field): DenoiseImage without -d on a 2D image: ANTs asks a new ImageIO for the dimension before it reads the header and NIfTI’s answers 3, so it denoises and writes a 3D image (ANTS-99, as Atropos); LARMORX used the file’s dimension in both modes. Now reproduce takes 3, the default the file’s dimension, both naming ANTS-99 (N4BiasFieldCorrection likewise)
  • fixed in LARMORX DenoiseImage (replica), 1 case(s): denoised: data type and shape, denoised: header (every field), noise: data type and shape, noise: header (every field): DenoiseImage without -d on a 2D or 4D image: ANTs takes it as 3D (ANTS-99); now reproduced, both modes naming it

Notices: a correction named in a case where it changed nothing the input shows (defect, cases): DenoiseImage: ANTS-43 (22), ITK-01 (4).

Every case is drawn from its tool, the run's seed and its number, and replays anywhere: how random testing works; each difference's triage: validation/difftest/triage.toml.

Known limits

The comparison with ANTs is made on Linux x86-64, against ANTs at one thread; ANTs at several threads has no single result to compare with. The cross-platform checks (how they work) hold the other five platforms to LARMORX’s own Linux output.

History

  • 2026-10-12: ANTS-43 corrected by default (register).
  • 2026-10-21: the same command run again by ANTs 2.6.5, six times at 8 threads and three times at 12: one result each (earlier runs at 8 threads differed, findings).
  • 2026-10-23: ANTS-44 decided and corrected by default.
  • 2026-10-26: the data panel’s record, sMRIPrep’s denoising of every T1w of the fourteen runs and four more images.

Data credits

The data this page shows or names: each dataset with its version, DOI and licence, from LARMORX's test-data catalog. Public-domain dedications (CC0, PDDL) ask nothing; the datasets are cited all the same.